Â鶹ÒùÔº

It's August, which means Hot Science Summer is two-thirds over. This week, NASA released an exceptionally pretty photo of Mars, a sharp panorama color altered to make the sky blue (???). California health authorities are warning hunters and trappers about contaminated game after one trapper caught a wild pig with bright blue muscle tissue. The pigs and other wildlife may have been exposed to the anticoagulant rodenticide diphacinone. And an international team of astronomers identified the oldest known black hole ever confirmed—the object was present 500 million years after the Big Bang.

Additionally, functional brain scanning suggests that video games are more engaging (in a good way) than television viewing or social media use; Chinese astronomers have teased out a triple black hole system from LIGO data; and AI company Anthropic is reporting a new method to prevent LLMs from becoming evil.

Black hole pair may have had a larger chaperone

Researchers from the Shanghai Astronomical Observatory report evidence that a recently observed binary black hole merger may have occurred within the gravitational field of a third object, quite possibly a supermassive black hole. This suggests an entirely new model for binary black hole mergers and the interactions of highly compact objects.

The team analyzed a gravitational wave event called GW190814, the merger of two black holes. The pair was an extreme example of a binary merger; the size ratio between the objects was 10:1. This suggests that they may once have been part of a triple system with a supermassive black hole and gradually spun together through gravitational interactions.

The researchers noted that if a binary black hole merger occurs within the gravity of a third object, the orbital motion around that third object would create a line-of-sight acceleration: to observers on Earth, this period of relative acceleration would create a Doppler alteration of the gravitational wave frequency. An analysis of the event showed that line-of-sight acceleration significantly outperformed a traditional "isolated binary" model in explaining the gravitational wave signal.

Study contradicts all dads everywhere

As governments around the world work toward internet age verification regimes and social media bans for young people, researchers at Swinburne University of Technology as they watched television, played video games or used phone-based social media apps. Employing functional near-infrared spectrography, they found that different forms of screen use had different patterns of brain activity.

Oxygenated hemoglobin levels increased more following social media use and gaming as compared to television viewing, and deoxygenated hemoglobin increased more following gaming. This suggests that interactive types of entertainment are more engaging than passive forms.

Dr. Alexandra Gaillard, one of the researchers, said, "Interestingly, though, when it came to social media, people reported feeling less focused—and those who felt less focused also showed lower levels of brain activity. On the flip side, gaming actually helped boost focus and showed a rise in deoxygenated hemoglobin, which means the brain was actively using more of the oxygen it was getting. In other words, gaming seemed to get the brain working harder in a good way."

Researchers identify and neutralize evil

AI company Anthropic, creators of the LLM Claude, report a to prevent artificial intelligence from becoming evil. The method identifies patterns of activity called persona vectors in neural networks. These patterns control the model's character traits. The researchers used two open-source LLMs to test removing and manipulating three persona vectors: evil, sycophancy and hallucination, describing the tendency of an LLM to make up information.

They call their method "steering"; when they manipulated the "evil" vector, the model would produce unethical responses; steering the "sycophancy" vector resulted in responses that flattered the user. Steering "hallucination" resulted in now-familiar, confidently nonfactual responses to prompts.

The authors write, "Our method for [steering] is somewhat counterintuitive: We actually steer the model toward undesirable persona vectors during training. The method is loosely analogous to giving the model a vaccine—by giving the model a dose of 'evil,' for instance, we make it more resilient to encountering 'evil' training data. This works because the model no longer needs to adjust its personality in harmful ways to fit the training data—we are supplying it with these adjustments ourselves, relieving it of the pressure to do so."

© 2025 Science X Network